Automatic Speech Recognition Framework for Indian Languages

In this project, we worked on developing several automatic speech recognition models for Indian languages, namely Tamil, Telugu and Gujarati using the Kaldi Speech Recognition Toolkit. A HMM-GMM acoustic model in conjugation with a n-gram language model was initially built for converting speech in the above mentioned languages to text. To obtain improved Word Error Rates, a Time Delay Neural Network (TDNN) was run. A Recurrent Neural Network based Language Model (RNNLM) pipeline was then set up to improve the contextual information compared to the n-gram language model. To achieve End-To-End speech recognition, CTC (Connectionist Temporal Classification) was used in conjugation with an Encoder-Decoder framework. A detailed ananlysis of this framework was performed to obtain best results.